Frequent frames as a cue for grammatical categories in child directed speechIntroductionPsycholinguistic evidence relevant to framesComparisons with previous distributional approachesExperiment 1MethodResultsDiscussionExperiment 2MethodResults and discussionGeneral discussionMechanisms for identifying informative framesErrorsFrom distributional to grammatical categoriesCross-linguistic applicabilityFrequent frames in acquisitionSummary and conclusionAcknowledgementsReferencesFrequent frames as a cue for grammaticalcategories in child directed speechToben H. Mintz*Departments of Psychology and Linguistics, and Program in Neuroscience,University of Southern California, Los Angeles, CA, USAReceived 24 March 2003; accepted 10 July 2003AbstractThis paper introduces the notion of frequent frames, distributional patterns based on co-occurrencepatterns of words in sentences, then investigates the usefulness of this information in grammaticalcategorization. A frame is defined as two jointly occurring words with one word intervening.Qualitative and quantitative results from distributional analyses of six different corpora of childdirected speech are presented in two experiments. In the analyses, words that were surrounded by thesame frequent frame were categorized together. The results show that frequent frames yield veryaccurate categories. Furthermore, evidence from behavioral studies suggests that infants and adultsare sensitive to frame-like units, and that adults use them to categorize words. This evidence, alongwith the success of frames in categorizing words, provides support for frames as a basis for theacquisition of grammatical categories.q 2003 Elsevier B.V. All rights reserved.Keywords: Language acquisition; Grammatical categories; Distributional analysis; Corpus analysis1. IntroductionGrammatical categories (e.g. noun verb, etc.) are fundamental building blocks ofgrammar, yet it is not fully known how child language learners initially categorize words.There has been recent interest in the idea that distributional information carried by the co-occurrence patterns of words in sentences could provide a great deal of informationrelevant to grammatical categories. For example, words in position X in sentences0022-2860/$ - see front matter q 2003 Elsevier B.V. All rights reserved.doi:10.1016/S0010-0277(03)00140-9Cognition 90 (2003) 91–117www.elsevier.com/locate/COGNIT*Department of Psychology, SGM 501, University of Southern California, Los Angeles, CA 90089-1061,USA. Tel.: þ 1-213-740-2253; fax: þ1-213-746-9082.E-mail address: [email protected] (T.H. Mintz).containing the English fragment in (1) are likely to belong to the same grammaticalcategory, verb.(1) … to X to …Building on early ideas from structural linguistics and related proposals for acquisition(Harris, 1951; Maratsos & Chalkley, 1980), several recent studies have investigatedwhether distributional patterns in English-learning children’s input could be a reliablesource of information for learning the category structure of the language (Cartwright &Brent, 1997; Mintz, Newport, & Bever, 1995, 2002; Redington, Chater, & Finch, 1998).These studies demonstrated that distributional patterns were informative and potentiallyviable bases for an initial categorization of words, and provided evidence againstspeculative yet influential claims that distributional information would be extremelyunreliable. As an example of the potential problems faced by distributional approaches,Pinker (1987) argued that, given sentences in (2a,b), a distributional learner wouldincorrectly categorize fish and rabbits together and, hearing (2c), would incorrectlyassume that (2d) is also permissible.(2) a. John ate fish.b. John ate rabbits.c. John can fish.d.pJohn can rabbits.The crux of the problem exemplified in (2) is that a given word form (in this case, fish) canbelong to multiple categories and thus occur in different syntactic contexts (e.g. as a nounin 2a or a verb in 2c), potentially providing misleading category information. Pinkerargued that the resulting erroneous generalizations would be common, and would render adistributional approach to categorization untenable.Another potential difficulty is that important distributional regularities are not alwayslocal, as in (1), but can occur over a variable distance, as in (3) (Chomsky, 1965; Pinker,1987).(3) … to hurriedly and effortfully Xto…Here, the informative verb environment for X in (1) (… to X to …) spans many words. Thefundamental issue is that lexical adjacency patterns are variable: in any particularutterance, the words in a specific position relative to a target is somewhat accidental, and alearner that categorized only from fixed positions could be led to make erroneousgeneralizations. Thus, another question is how the learner is to know which environmentsare important and which should be ignored. Distributional analyses that consider allthe possible relations among words in a corpus of sentences would be computationallyunmanageable at best, and impossible at worst.Hence, it was an important advance when results from recent empirical investigationsinto the viability of distributional approaches to categorization revealed that, in children’sactual input, these potential problems do not significantly undermine the informativenessT.H. Mintz / Cognition 90 (2003) 91–11792of distributional patterns. For example, Mintz et al. (1995, 2002) and Redington et al.(1998) showed that local contexts restricted to immediately adjacent words can beinformative as to a word’s category membership. Such findings suggested that, althoughproblematic environments may exist, there is nonetheless enough “signal” in thedistributional patterns compared to the noise created by the problematic environments thatcategorization from distributional patterns is not intractable. Moreover, by showing thatlocal contexts are informative, these findings suggested a solution to the problem of therebeing too many possible environments to keep track of: focusing on local contexts mightbe sufficient.Although prior studies have shown distributional information to be useful forcategorizing words, many open questions remain and much research is still needed todetermine what type of distributional information is especially informative, and whatkinds of distributional cues infants and young children are sensitive to and use incategorizing words. Questions also remain as to how distributionally defined